Webinar: Amira Abbas, The feasibility of quantum backpropagation

The success of modern deep learning hinges on the ability to train neural networks at scale. Through clever reuse of intermediate information, backpropagation facilitates training through gradient computation at a total cost roughly proportional to running the function, rather than incurring an additional factor proportional to the number of parameters – which can now be in the trillions. This motivates interest in determining whether parameterized quantum models, such as those used in variational algorithms, can too, train as efficiently with gradient-based methods.

In this talk, I will discuss why this task is difficult in a quantum setting, where all known gradient methods fail to achieve backpropagation-scaling, unless special case models are considered. Unfortunately, in general, reusing information (similar to how classical backpropagation works) will not succeed either and we demonstrate this failure through connections to gentle measurement. These no-go results could dramatically change the landscape for quantum machine learning models, as they highlight additional difficulties that variational approaches will face at scale, unless alternative optimization methods or quantum models are found.